Agent Skills

Evidence is limited to README, LICENSE, and two GitHub Actions workflows; no SKILL.md bodies are provided, so runtime permission scope, confirmation flows, data-flow disclosure, and sensitive-data handling cannot be verified. least_privilege scores 1: workflows explicitly declare permissions: contents: read, a good least-privilege signal, but this covers only CI, not the skills themselves. user_confirmation scores 1: README mentions tidy is report-only by default and autoship has a release flow, hinting at confirmation steps, but no implementation evidence. data_flow_transparency scores 1: rebuild-web.yml comments explain the Vercel deploy hook and CDN cache behavior in detail, which is partial transparency, but skill-level data flows are undisclosed. sensitive_data_handling scores 1: validator tests show a no-absolute-paths check, indicating awareness of path leakage, but no sensitive-data classification or redaction policy. dependency_security scores 1: workflows use mainstream actions (actions/checkout@v4, ruby/setup-ruby@v1, actions/setup-python@v5) but do not pin SHAs, and skill dependencies (npx skills, Ruby, Python packages) are neither listed nor locked. external_effects scores 1: autoship publishes to npm and pr-babysitter modifies PRs, which are external side effects, but no rollback or blast-radius description. rollback scores 1: workflows have concurrency and retries, but no skill-level rollback mechanism is described. source_attribution scores 2: LICENSE clearly attributes Matthew Blode and README names the author and source, so attribution is clear; however, the publisher is not verified by FollowAgents, so no higher score is given.

self_consistency scores 2: tests in maintenance/tests assert consistency between the ghostwriter Surfaces table and references/surfaces.md and exercise the validator CLI contract, showing internal consistency checks; however, the test files have structural issues (test_validator.py has test methods after the if __name__ block that will not execute), weakening reliability. dependency_availability scores 1: workflows depend on ruby 3.3, python 3.12, npx skills, and other external tools, but no version locking or availability guarantees are provided. failure_messages scores 2: rebuild-web.yml emits clear errors and exit codes for missing secrets, HTTP 403/429, and timeouts, and validator tests verify FAIL row formats, so failure messaging is reasonably clear.

audience_and_scenarios scores 2: README categorizes 20+ skills under Architecture, Design, Writing, Quality, Shipping, and Authoring, covering broad scenarios for coding agents and human users. capability_boundaries scores 1: boundaries can only be inferred from one-line descriptions; no SKILL.md body states what each skill does not do or when it is inapplicable. trigger_precision scores 1: README notes app-verification is called by name and test-audit is invoked explicitly, hinting at some trigger conditions, but there is no systematic trigger-word or exclusion documentation. environment_fit scores 1: the install command targets codex and claude-code, but other agent environments, operating systems, and Node version requirements are unspecified.

information_architecture scores 2: README is well structured, organizing skills by category with links to each SKILL.md, and clearly separates install and license sections. install_notes scores 2: provides a concrete npx skills add command with -g, --agent, and -y flags, making installation actionable. naming_stability scores 1: skill names follow a consistent kebab-case style, but there is no rename policy or deprecated-name record (tests reference retired-names.tsv, but the file is not provided). examples_and_faq scores 1: README lists only one-line descriptions with no usage examples or FAQ. known_limitations scores 1: rebuild-web.yml comments disclose operational limits such as CDN caching and 403 rate limiting, but skill-level known limitations are not stated. license scores 3: LICENSE.md contains the full MIT text, and README and badge both mark MIT, with clear copyright year and author. versioning_changelog scores 0: no CHANGELOG, no version numbers, no release history. maintenance_responsibility scores 1: README names the author and links to blode.co, but there is no maintenance commitment, issue-response expectation, or security contact.

output_usability scores 1: skill output formats and usability cannot be judged from the available files; README only describes functionality. marginal_value scores 1: skills cover UI audits, typography, docs, PR review, and releases, which appear useful, but there is no comparison with existing tools or evidence of effect. cost_benefit scores 1: installation is a single npx command, so cost is low; however, runtime token consumption, time cost, and benefit are unsupported by data.

claim_traceability scores 1: README claims 78 typography rules, 27 ax rules, 38 dx rules, etc., but provides no rule list or verifiable source. cross_source_corroboration scores 1: README and test files partially corroborate each other on the ghostwriter Surfaces table and validator contract, but most skill claims have no second source. fact_inference_separation scores 1: README mixes functional descriptions with marketing copy (e.g., 'Nobody ships AI slop on purpose') without separating fact from assertion.